French ornithologists have proposed to automate the collection of data for the recognition algorithms with the help of RFID. Method they tested on the birds at their feeders placed RFID sensors that recorded the presence of birds and sent the signal to the camera running on a Raspberry Pi. The algorithm based on convolutional neural networks trained on the collected data, have learned to recognize the birds on the back with accuracy above 90 percent. Detailed operation of the system described in the article published in Methods in Ecology and Evolution.
Modern methods of recognition of objects on images (e.g., using convolutional neural networks) are well suited and has long been used for identifying species and counting individuals in the wild. To study animals in captivity they are used too: in addition to the classic recognition and counting — and also in order to follow individuals (as, for example, PrimNet, presented two years ago). The solution to this problem, however, is slightly more complicated, mainly because requires this quantity of data, which, on the one hand, enough for effective learning algorithm, on the other hand, will be sufficient to ensure that an individual can be recognized at any angle.
Andre Ferreira (André Ferreira) from the University of Montpellier proposed to automate the process of collecting images of individuals for further learning algorithms. Scientists have focused on birds: the populations of common social weavers (Philetairus socius) and great Tits (Parus major) living in the reserve and populations of Zebra finches (Taeniopygia guttata) living in captivity.
For protected birds and birds living in captivity, the data collection differed slightly. The first birds were born marked with RFID-tags, therefore they placed on the bird feeders receivers register the individual, sending a signal to the camera connected to the Raspberry Pi and the camera took the picture. Image weavers, scientists collected two weeks, and pictures of the great tit is week. Birds in captivity, not labeled, photographed in individual cells-the feeders for four hours: the camera connected to the Raspberry Pi took pictures individuals every two seconds. All cameras were set on the opposite side of the trough — to dataset only got pictures of the bird’s backs.